Interesting critique of British education by outgoing advisor (warning: some politics)

The soon-to-be-resigning Dominic Cummings, advisor to the Education Secretary of the Coalition government, has released a 250-page manifesto describing the problems of the British educational establishment (“the blob” in Whitehall parlance) and offering solutions. I post this here because both his analysis and recommendations are likely to be interesting to LW, in particular an increased emphasis on STEM, broader knowledge of the limits of human reasoning and how they relate to managing complex systems, an appreciation for “agenty”-ness in organizational leadership, whole-brain emulation, intelligence enhancement, recursive self-improving AGI, analysis of human interactions on a firm evolutionary-psychological basis, and a rejection of fashionable pseudoscientific theories of psychology and society. Relevant extracts:

This essay is aimed mainly at ~15-25 year-olds and those interested in more ambitious education and training for them. Not only are most of them forced into mediocre education but they are also then forced into dysfunctional institutions where many face awful choices: either conform to the patterns set by middle-aged mediocrities (don’t pursue excellence, don’t challenge bosses’ errors, and so on) or soon be despised and unemployed. Some of the ideas sketched here may help some to shape their own education and create their own institutions. As people such as Linus Torvald (Linux) or Mark Zuckerberg (Facebook) have shown, the young are capable of much more than the powerful and middle-aged, who control so many institutions, like to admit.2 A significant change in education and training could also help us partly ameliorate the grim cycle of predictable political dysfunction.

Although it is generally psychologically unpleasant to focus on our problems and admit the great weaknesses of our institutions, including what we in politics and our leaders do not understand, it is only through honest analysis that progress is possible. As Maxwell said, ‘Thoroughly conscious ignorance is a prelude to every real advance in knowledge.’ Reliable knowledge about what works (how do people learn, what teaching methods work, how to use technology) should be built cumulatively on the foundation of empirical tests as suggested by physicist Carl Wieman and others (cf. Section 6 and Endnote). New systems and curricula would work in different ways for a school, a University, or an equivalent to a boxing gym for politicians if such a thing existed.

In particular, it is suggested that we need an ‘Odyssean’ education so that a substantial fraction of teenagers, students and adults might understand something of our biggest intellectual and practical problems, and be trained to take effective action.

The Nobel-winning physicist, Murray Gell Mann, one of the architects of the Standard Model of particle physics and namer of the ‘quark’,3 has described a scientific and political need for an ‘Odyssean’ philosophy that can synthesise a) maths and the natural sciences, b) the social sciences, and c) the humanities and arts, into necessarily crude, trans-disciplinary, integrative thinking about complex systems.

’Today the network of relationships linking the human race to itself and to the rest of the biosphere is so complex that all aspects affect all others to an extraordinary degree. Someone should be studying the whole system, however crudely that has to be done, because no gluing together of partial studies of a complex nonlinear system can give a good idea of the behavior of the whole...

’Those who study complex adaptive systems are beginning to find some general principles that underlie all such systems, and seeking out those principles requires intensive discussions and collaborations among specialists in a great many fields. Of course the careful and inspired study of each specialty remains as vital as ever. But integration of those specialities is urgently needed as well. Important contributions are made by the handful of scholars and scientists who are transforming themselves from specialists into students of simplicity and complexity or of complex adaptive systems in general…

‘[There is] the distinction (made famous by Nietzsche) between “Apollonians”, who favor logic, the analytical approach, and a dispassionate weighing of the evidence, and “Dionysians”’, who lean more toward intuition, synthesis, and passion…[4] But some of us seem to belong to another category: the “Odysseans”, who combine the two predilections in their quest for connections among ideas… We need to celebrate the contribution of those who dare take what I call “a crude look at the whole”…

’… broadly integrative thinking is relegated to cocktail parties. In academic life, in bureaucracies, and elsewhere, the task of integration is insufficiently respected. Yet anyone at the top of an organization … has to make decisions as if all aspects of a situation, along with the interaction among those aspects, were being taken into account. Is it reasonable for the leader, reaching down into the organization for help, to encounter specialists and for integrative thinking to take place only when he or she makes the final intuitive judgements?

’[A] multiplicity of crude but integrative policy studies, involving not just linear projection but evolution and highly nonlinear simulation and gaming, may provide some modest help in generating a collective foresight function for the human race…

’Given the immense complexity of the numerous interlocking issues facing humanity, foresight demands the ability to identify and gather great quantities of relevant information; the ability to catch glimpses, using that information, of the choices offered by the branching alternative histories of the future, and the wisdom to select simplifications and approximations that do not sacrifice the representation of critical qualitative issues, especially issues of values…

’Computers … can serve us both by learning or adapting themselves and by modelling or simulating systems in the real world that learn or adapt or evolve. .. Powerful computers are essential for assistance in looking into the future, but we must not allow their use to bias the formulation of problems toward the quantifiable and analyzable at the expense of the important.’5

One of the things that ‘synthesizers’ need to learn is the way that many themes cut across subjects and generate new subjects. Trans-disciplinary studies of complex systems have been profoundly affected by connected intellectual revolutions in physics, maths, logic, computation, and biology, though these connections have so far barely been integrated in school or university curricula. Ideas about ‘information’ cut across physics (thermodynamics and entropy),7 computation (bits, qubits, and ‘artificial agents’), economics (‘economic agents’, Hayek’s description of prices as an ‘information discovery process’), evolution (genetic networks),8 the brain (neuronal networks), ‘intelligence failures’9 and other subjects. Physics is inseparable from old philosophical questions (it is ‘experimental metaphysics’) and it is merging with computer science to produce quantum computation and ‘quantum information theory’.10 Evolutionary ideas took hold in economics (Hume, Smith) and biology (Darwin), and they have now been incorporated into computer science (e.g. ‘genetic algorithms’) in order to ‘evolve’ solutions to problems in large search spaces, and they have suggested ideas for engineering solutions. (For example, it is suggested that dangers such as bioterrorism or pandemics should be defended by developing ‘artificial immune systems’ in which defences operate according to the evolutionary principles of i) generating lots of solutions with random variation, and ii) differential replication of the most effective agents, instead of reliance on traditional centralised institutions that make similar mistakes repeatedly.) Machine intelligence (or ‘automated reasoning’) has been shaped by, and in turn is reshaping, ideas about the mind and is now used to design computers, including those that are used to investigate the mind. Behavioural genetics, evolutionary psychology, cognitive science and neuroscience are reshaping not only economics (e.g. fMRI scans of people playing financial games) and history (e.g. the influence of evolved antipathy for out-groups) but also how we design institutions (e.g. how we evolved to succeed in the sexual politics of small, primitive collectivist tribes hence many of the standard features of internal politics). As will be discussed, there are various developments in education that seek to reflect these trans-disciplinary themes: e.g. the Nobel-winning neuroscientist, Eric Kandel, plans a new PhD programme combining neuroscience, psychology and art history as part of the ‘Mind, Brain, Behaviour Project’ at Columbia.11

Neuroscience, cognitive science, behavioural genetics, and evolutionary biology have developed our understanding of the mind. They and other disciplines have combined with Moore’s Law and scanning technology to provide increasingly accurate maps and quantitative models of the brain107 (which Obama promised federal support for in 2013) and rapidly improving brain-computer interfaces.108 We can look inside our minds with tools that our minds have given us and watch the brain watching itself. These developments have undermined the basis for Descartes’ Ghost in the Machine, Locke’s Blank Slate, and Rousseau’s Noble Savage (pace the current, sometimes misguided, backlash).109

The brain consists of ~80 billion (8x1010) neurons and ~100 trillion (1014) synapses. Operating on ~20 watts it performs ~1017 floating point computations per second. As well as the brain-computer interfaces already underway, various projects plan to map the brain completely (e.g the Connectome Project), simulate the brain (e.g Markram’s project), and build new computer architectures modeled on the brain and performing similarly to the brain.

For example, the most successful government technology developer, DARPA, has made robotics and machine intelligence a priority with projects such as the SyNAPSE Project (led by IBM’s Modha) to create ‘a brain inspired electronic “chip” that mimics that function, size, and power consumption of a biological cortex.’110 They have recently announced progress in building this new architecture, fundamentally different to the standard ‘von Neumann architecture’ of all normal computers.111

The Human Brain Project is aiming to model a human brain and in 2012 was awarded €1 billion by the EU. In 2005, a single neuron was simulated; in 2008, a cortical column (104 neurons); in 2011, 100 columns (106 neurons). Markram plans for a full rodent brain simulation in 2014 and a human brain simulation 2020-25.

However, so far SyNAPSE and The Human Brain Project have not demonstrated how simulations connect to observable behaviours. In November 2012, Eliasmith et al (Science, 30/​12/​2012) published details of a large-scale computational model of the brain (Semantic Pointer Architecture Unified Network, or ‘Spaun’) intended to bridge ‘the brain-behaviour gap’ by simulating complex behaviour. Spaun is a ‘spiking neuron model’ of 2.5m simulated neurons organised into subsystems resembling different brain areas. All inputs are images of characters; all outputs are movements of a physically modeled arm. Incoming visual images are compressed by successive layers of the network that extract increasingly complex information, and simple internally generated commands (e.g. ‘draw a number’) are ‘expanded’ into complex mechanical movements. Spaun simulates the working memory and an ‘action selection system’. It performs eight tasks such as image recognition, copy drawing, reinforcement learning, and a fluid reasoning task ‘isomorphic to the induction problems from the Raven’s Progressive Matrices (RPM) test for fluid intelligence’. Spaun managed to pass some basic aspects of an IQ test. Spaun is not task-specific so the model could be extended to other tasks and scaled-up in other ways.

In 2013, Alex Wissner-Gross, a Harvard computer scientist, published a paper in Physical Review in ‘an attempt to describe intelligence as a fundamentally thermodynamic process’, proposing that intelligence can spontaneously emerge from the attempt to maximise freedom of action in the future. He built a software programme, ‘ENTROPICA’, designed to maximise the production of long-term entropy of any system it finds itself in. ENTROPICA then solved various problems including intelligence tests, playing games, social cooperation, trading financial instruments, and ‘balancing’ a physical system and so on.

‘We were actually able to successfully reproduce standard intelligence tests and other cognitive behaviors, all without assigning any explicit goals… ’

Think of games like chess or Go in which good players try to preserve as much freedom of action as possible. When the best computer programs play Go, they rely on a principle in which the best move is the one which preserves the greatest fraction of possible wins. When computers are equipped with this simple strategy—along with some pruning for efficiency—they begin to approach the level of Go grandmasters…

’Our causal entropy maximization theory predicts that AIs may be fundamentally antithetical to being boxed. If intelligence is a phenomenon that spontaneously emerges through causal entropy maximization, then it might mean that you could effectively reframe the entire definition of Artificial General Intelligence to be a physical effect resulting from a process that tries to avoid being boxed...

’The conventional storyline has been that we would first build a really intelligent machine, and then it would spontaneously decide to take over the world…We may have gotten the order of dependence all wrong. Intelligence and superintelligence may actually emerge from the effort of trying to take control of the world—and specifically, all possible futures—rather than taking control of the world being a behavior that spontaneously emerges from having superhuman machine intelligence…

’The recursive self-improving of an AI can be seen as implicitly inducing a flow over the entire space of possible AI programs. In that context, if you look at that flow over AI program space, it is conceivable that causal entropy maximization might represent a fixed point and that a recursively self-improving AI will tend to self-modify so as to do a better and better job of maximizing its future possibilities.

‘In the problem solving example, I show that cooperation can emerge as a means for the systems to maximize their causal entropy, so it doesn’t always have to be competition. If more future possibilities are gained through cooperation rather than competition, then cooperation by itself should spontaneously emerge, speaking to the potential for friendliness.’ (Interview.)

The education of the majority even in rich countries is between awful and mediocre. A tiny number, less than 1 percent, are educated in the basics of how the ‘unreasonable effectiveness of mathematics’ provides the ‘language of nature’ and a foundation for our scientific civilisation116 and only a small subset of that <1% then study trans-disciplinary issues concerning the understanding, prediction and control of complex nonlinear systems. Unavoidably, the level of one’s mathematical understanding imposes limits on the depth to which one can explore many subjects. For example, it is impossible to follow academic debates about IQ unless one knows roughly what ‘normal distribution’ and ‘standard deviation’ mean, and many political decisions, concerning issues such as risk, cannot be wisely taken without at least knowing of the existence of mathematical tools such as conditional probability. Only a few aspects of this problem will be mentioned.

There is widespread dishonesty about standards in English schools,117 low aspiration even for the brightest children,118 and a common view that only a small fraction of the population, a subset of the most able, should be given a reasonably advanced mathematical and scientific education, while many other able pupils leave school with little more than basic numeracy and some scattered, soon-forgotten facts. A reasonable overall conclusion from international comparisons, many studies, and how universities have behaved, is that overall standards have roughly stagnated over the past thirty years (at best), there are fewer awful schools, the sharp rises in GCSE results reflect easier exams rather than real educational improvements, and the skills expected of the top 20 percent of the ability range studying core A Level subjects significantly declined (while private schools continued to teach beyond A Levels), hence private schools have continued to dominate Oxbridge entry while even the best universities have had to change degree courses substantially

There is hostility to treating education as a field for objective scientific research to identify what different methods and resources might achieve for different sorts of pupils. The quality of much education research is poor. Randomised control trials (RCTs) are rarely used to evaluate programmes costing huge amounts of money. They were resisted by the medical community for decades (‘don’t challenge my expertise with data’)119 and this attitude still pervades education. There are many ‘studies’ that one cannot rely on and which have not been replicated. Methods are often based on technological constraints of centuries ago, such as lectures. Square wheels are repeatedly reinvented despite the availability of exceptional materials and subject experts are routinely ignored by professional ‘educationalists’.120 There is approximately zero connection between a) debates in Westminster and the media about education and b) relevant science, and little desire to make such connections or build the systems necessary; almost everybody prefers the current approach despite occasional talk of ‘evidence-based policy’ (this problem is one of the reasons we asked Ben Goldacre to review the DfE’s analysis division). The political implications of discussing the effects of evolutionary influences on the variance of various characteristics (such as intelligence (‘g’) and conscientiousness) and the gaps between work done by natural scientists and much ‘social science’ commentary have also prevented rational public discussion (cf. Endnote on IQ).121

Westminster and Whitehall have distorted incentives to learn and improve,122 have simultaneously taken control of curricula and exams and undermined the credibility of both, and have then blamed universities for the failures of state schools123 and put enormous pressure on Universities and academics not to speak publicly about problems with exams, which has made rational discussion of exams impossible. Most people with power in the education system are more worried about being accused of ‘elitism’ (and ‘dividing children into sheep and goats’) than they are about problems caused by poor teaching and exams and they would rather live with those problems than deal with those accusations.124

[124 E.g. Almost everybody the DfE consulted 2011-13 about curriculum and exam reform was much more concerned about accusations of elitism than about the lack of ambition for the top 20%. Although they would not put it like this, most prominent people in the education world tacitly accept that failing to develop the talents of the most able is a price worth paying to be able to pose as defenders of ‘equality’. The insistence that ~95% of pupils be able to take the same exam at 16 means (if one assumes symmetrical exclusions) that the exam must embrace plus and minus two standard deviations on the cognitive ability range: i.e. they exclude only the bottom 2.5% (i.e. an IQ of <~70) and top 2.5% (i.e an IQ of >~130, which is the average Physics PhD).]

There is huge variation in school performance (on exams that are sub-optimal) among schools with the poorest children. About a quarter of primaries have over a quarter of their pupils leave each year who are not properly prepared for basic secondary studies and few such pupils enjoy a turnaround at secondary;125 other primaries (including in the poorest areas) have <5% in such a desperate situation. Consider a basic benchmark: getting four-fifths of pupils to at least a ‘C’ in existing English and Maths GCSE. A small minority of state schools achieve this, while others with similar funding and similarly impoverished pupils struggle to get two-fifths to this level. This wide variety in performance combined with severe limits on direct parent choice means the system is partly rationed by house price.126

This wide variety in performance also strongly suggests that the block to achieving this basic benchmark is the management and quality of teaching in the school; the block is not poverty,127 IQ, money,128 lack of innovation, or a lack of understanding about how to teach basics. Making a transition to a school system in which ~4/​5 meet this basic level is therefore an issue of doing things we already know how to do; the obstacles are political and bureaucratic (such as replacing management and bad teachers despite political resistance and legal complexity), although this must not blind us to the fact that most variation in performance is due to within school factors (including genetics) rather than between school factors (see below).

There are various problems with maths and science education…

The Royal Society estimates that ~300,000 per year need some sort of post-GCSE Maths course but only ~100,000 do one now. About 610 now get at least a C in English and Maths GCSE; most never do any more maths after GCSE.129 There is no widely respected ‘maths for non-maths specialists’ 16-18 course (see below).130 About 70-80,000 (~1/​10 of the cohort) do Maths A Level each year (of these ~⅓ come from private schools and grammars)131 and ~1-2% also do Further Maths. In the last year for which we have data, ~0.5% (3,580 pupils) went on to get A or A* in each of A Level Maths, Further Maths, and Physics.132 Further, many universities only demand GCSE Maths as a condition of entry even for scientific degrees, so ~20% of HE Engineering entrants, ~40% of Chemistry and Economics entrants, and ~60-70% of Biology and Computer Science entrants do not have A Level Maths. Less than10% of undergraduate bioscience degree courses demand A Level Maths.

Because of how courses have been devised, ~4/​5 pupils leave England’s schools without basic knowledge of subjects like logarithms and exponential functions which are fundamental to many theoretical and practical problems (such as compound interest and interpreting a simple chart on a log scale), and unaware of the maths and physics of Newton (basic calculus and mechanics). Less than one in ten has a grasp of the maths of probability developed in the 19th Century such as ‘normal distributions’ and the Central Limit Theorem (‘bell curves’) and conditional probability.133 Only the 1-2% doing Further Maths study complex numbers, matrices and basic linear algebra. Basic logic and set theory (developed c. 1850-1940) do not feature in Maths or Further Maths A levels, so almost nobody leaves school with even a vague idea of the modern axiomatic approach to maths unless they go to a very unusual school or teach themselves.

133 Gigerenzer’s ‘Reckoning With Risk’ has terrifying stats on the inability of trained medical professionals making life and death decisions to understand the basics of conditional probability, which is not covered in the pre-16 curriculum (cf. Endnote). Current A Level modules have conditional probability and normal distributions in S1 and S2 (not compulsory), so one could have an A* in A Level Maths and Further Maths without knowing what these are. Data on who does which modules is not published by exam boards.

(...)

The education world generally resists fiercely the idea that a large fraction of children can or should be introduced to advanced ideas but we could substantially raise expectations without embracing ‘Ender’s Game’. It is almost never asked: how could we explore rigorously how ambitious it is realistic to be? If you ask, including in the Royal Society, ‘what proportion of kids with an IQ of X could master integration given a great teacher?’, you will get only blank looks and ‘I don’t think anyone has researched that’. Given the lack of empirical research into what pupils with different levels of cognitive ability are capable of with good teachers, research that obviously should be undertaken, and given excellent schools (private or state) show high performance is possible, it is important to err on the side of over-ambition rather than continue the current low expectations. Programmes in America have shown that ‘adolescents scoring 500 or higher on SAT-M or SAT-V by age 13 (top 1 in 200), can assimilate a full high school course (e.g., chemistry, English, and mathematics) in three weeks at summer residential programs for intellectually precocious youth; yet, those scoring 700 or more (top 1 in 10,000), can assimilate at least twice this amount…’ (Lubinski, 2010). (See Endnote.)